About a year ago FiveThirtyEight published a short article called “Is Your D&D Character Rare?”. It was a product of a deal between Curse and FiveThirtyEight which meant the data was not available to anyone else. I was a little jealous that I couldn’t play with the data and disappointed that they only counted class race combinations and called it a day.
Shortly after, I released a few add-ons (1,2) for a popular mobile application which allowed me to collect my users’ character sheets.
After 3.5 months of data collection I have a whopping… 646 unique characters in my database that I can play with. Well… I’m not as popular as DnDBeyond but I don’t see anyone else waving around thousands of character sheets for us to data mine, so it’ll have to do.
To start with let’s redo the table from FiveThirtyEight. I am not going to pretend like I have many thousands of samples so instead of per 100,000 this shows class and race combinations per 100 characters. In FiveThirtyEight’s table, characters with multiple classes count once for each class. Here I divided multiclassed characters based on the proportion of their class levels. For instance, a character who is a Fighter 5/Rogue 15 will add 0.75 to the rogue count and 0.25 to the fighter count. It also includes all races. Homebrew and UA classes are removed.
Despite the methodological differences, these results seem to correlate well (Spearman’s ρ=0.74) with DnDBeyond data even though we seem to disagree on what most popular classes and races are. Graph below shows the % occurrence of a class/race combination in FiveThityEight data and my data.
This is a little hard to visualize in a single plot. Alas we are short on space so you’re going to have to mouse over to see the details. Each colored section shows a character archetype’s proportion to the rest of the archetypes for the class. They are ordered from bottom to top in order of frequency so the brown always show the most popular archetype and it goes downhill (but upwards in the plot) from there.
I will never understand what people find in Champion but each to their own…
Jeremy Crawford once tweeted
Another piece of D&D data: a majority of D&D characters don’t use feats. Many players love the customization possible with feats, but a larger group of players is happy to make characters without feats. Feats are, therefore, not a driving force behind many players’ choices.
We can see whether or not our data agrees. On a surface look 33% of all characters have at least one feat. However, this is partially caused by the fact that a significant portion (49%) of our characters are between levels 1-3 and unless they are variant humans, they cannot have feats. We can see that by higher levels, feat adoption rates increase significantly, suggesting that once given the opportunity, players are likely to pick a feat.
It can be postulated players spend most of their time between levels 4-15. 52% of all characters in this range has at least one feat. As I later discovered, this also somewhat correlates with the data in DnDBeyond though the percentages here are higher overall.
Next logical step is to examine which classes picks which feats, and which feats are the most popular. The graph below shows which feat is selected the most and by which class. Multiclassed characters are merged into their own category to reduce clutter. Any feat that is selected only once is removed. Again, mouse over the bars to see details.
It is surprising that Elven Accuracy, a feat that is added in a supplement and resrticted to elves, is as popular as many core book feats that are known to be highly effective. 10% of all elves and half-elves have this feat. Its appeal to both ranged weapon attackers and casters seems to make it a good choice for elves from many walks of life. Another interesting bit is that the Magic Initiate feat seems be very popular amongst classes with spellcasting ability. I was always under the impression that Magic Initiate’s main use case would be to add some magic to a mundane class.
We can also look into how feats synergize with each other. The network below shows how often feats are selected together. Unique connections are removed. Node sizes represent how many times a feat appeared together with another feat. The thickness of the lines between the nodes are determined by the number of characters both feats appear in.
Before I say anything I have to declare the connections in this graph aren’t particularly strong. There are too many feats and I have too few characters for high number of feats to appear together. The strongest link in this graph is based on 4 observations.
Yet, as it stands, the connections seem quite intuitive so we are probably not staring at noise here. Robustnessof elven accuracy is visible in this graph as it is both selected by characters trying to optimize their ranged and spell attacks. Crossbow Expert-Sharpshooter is known to be an effective combination to boost damage. Sentinel-Polearm Master is anazing for battlefield control.
Since our dataset includes multiclassed characters, we can see which classed tend to appear together. Note that our sample size much smaller here (81 characters). Node sizes in the network below show how many times a class appeared in all multiclassed characters. The thickness of the lines between the nodes are determined by the number of characters both classes appear in. For instance, we see that most rangers multiclass with rogues, while most rogues multiclass with fighters.
While this network is good to show which classes tend to be chosen together, it doesn’t give much information about how classes are distributed. In the below graph we look at what is ratio of class levels in individual characters. A Fighter 5/Rogue 15 would appear as a 25% data point in the Fighter column and 75% in the Rogue column. This will give us information about which classes are dipped in and which ones are used as the main class.
While there is a high amount of variation in the data, some conventional wisdom pops up through the means. Warlock is famous for its dipping potential and a Cleric level synergizes nicely with many other class features. I am a proud player of a Cleric dipped Fighter myself. I would avoid reading too much into this though. The variance is too high and sample size is too low to make reliable inferences.
And finally let’s see which classes tend to appear in multiclassed builds compared to single classed ones
Ok that title is a stretch but we have a format to stick to.
Both multiclassing and picking feats are somewhat advanced character building rules. While making the character building process complicated, they can be used to create frighteningly affective combinations (or get stuck waiting till the end of the campaign till their build gets everything they want). Intuitively, it wouldn’t be surprising to see that multiclassers are more likely to get feats to optimize their builds. Indeed, we see that 59% of multiclassed characters above level 3 chose to get a feat as opposed to 52% of single classed counterparts. A modest yet statistically significant difference (p=6.6e-07).
My applications are they are purely utilitarian. One gives you a character sheet, the other is an interactive character sheet that automates your dice roll. It is somewhat reasonable to think that most people would be using them shortly before or during a game. Graphs below how many characters were created in each day of the week and below that there’s a punch card that shows individual hours.
Frankly not much to be said here. Most popular days of the week are obviously weekends and Friday. DnD takes time. More work = less DnD. Hours of day are somewhat unreliable as I didn’t correct for user time zones. US alone, which seems to be where most my users are coming from, can have 3 hours of difference. I could use IPs and detect locations to fix times but not going into that rabbit hole… How long before the game a player may want their character sheet is also a great source of variability. I mostly did this because I like punch cards…
Unique characters are acquired by grouping the characters that share the same name and class and picking the higher level version. This could have merged independent characters with tropey names like Grognak the Barbarian of Drizzt the Ranger but manual examination of the data showed no cases of characters who appear to be made by different people but still has the same name and class.
If a multiclassed character shares name with a single classed character, I assume they are duplicates if the single classed character is lower level and its class matches with one of the classes of the multiclassed character.
Any character above level 20 (there were 3) were removed.
8 Revised Rangers were merged back into the ranger class.
Most percentages are rounded to the nearest integer
As all data, this data comes with caveats. It is a subset of all DnD players who are using a particular mobile application who also know about and use my applications and consented to let me to keep their character sheets. I don’t have reason to think that these would be enriching certain character building choices but it’s something to keep in mind.
Here’s its source code. It’s not pretty.